Ice and snow tourism creative tour guiding and abnormity early warning system based on behavior state monitoring

Abstract

The invention discloses an ice and snow tourism cultural and creative guide and abnormity early warning system based on behavior state monitoring, and relates to the technical field of intelligent safety early warning. Comprising the following steps: acquiring multi-modal data such as environment voiceprint characteristics, user behavior states and user intentions through a data acquisition module; the multi-modal decision module then identifies external scene events based on the environmental voiceprint, identifies user intrinsic state identifications based on user behavior states, and associates the two to determine a decision context. When a safety related event is recognized and the user is in a high-risk state, the system judges to enter a safety priority mode, adaptively filters or adjusts a voice instruction of the user, ignores a non-safety request and preferentially generates an early warning trigger instruction; and the output module outputs corresponding prompt information. According to the invention, the navigation service and the safety early warning are intelligently coordinated in the complex ice and snow tourism scene, so that the safety of the user and the man-machine interaction intelligent level of the system are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent safety early warning technology, specifically to a snow and ice tourism cultural and creative product guidance and anomaly early warning system based on behavioral status monitoring. Background Technology

[0002] Ice and snow tourism, as an emerging cultural and tourism industry, has gained widespread popularity in recent years. To enhance the visitor experience, intelligent tour guide systems are widely used in scenic areas. Their functions have gradually expanded from basic route navigation and explanation of cultural and creative information to safety warnings, becoming an important component of smart scenic area construction.

[0003] Existing technologies have fundamental limitations in their early warning logic when dealing with the dynamic and complex scenario of ice and snow tourism. First, existing systems generally employ isolated judgment modes such as "environmental risk broadcasting" or "user behavior exceeding boundaries," failing to deeply correlate and collaboratively analyze real-time risks in the external environment with the user's own behavioral state. For example, when the system detects an acoustic event that may cause ice and snow collapse in the distance, it will issue a warning to all tourists, but it cannot determine which tourist is in a high-risk posture—facing away from the sound source and focused on taking photos—thus failing to achieve accurate delivery of warning information and prioritized intervention. Second, the interaction logic of existing tour guide systems is egalitarian and passive; that is, cultural and creative explanations, route planning, and safety warnings are given the same priority. The system cannot proactively interrupt non-critical service requests in critical moments. When a tourist listening to a historical story is simultaneously in danger, the system may delay safety alerts due to being busy outputting cultural and creative content. This lack of dynamic management capabilities in resource allocation and instruction priority at critical moments may lead to warning information being overwhelmed and the best intervention opportunity being missed. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a system for guiding and providing early warning of anomalies in ice and snow tourism and cultural and creative products based on behavioral status monitoring.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] This invention discloses a guided tour and anomaly early warning system for ice and snow tourism cultural and creative products based on behavioral state monitoring, including:

[0007] The data acquisition module is used to acquire environmental soundprint feature data, user behavior status data, and user intent data.

[0008] The multimodal decision module is used to execute:

[0009] External scene events are identified based on the environmental soundprint feature data, and the user's internal state identifier is identified based on the user behavior state data.

[0010] The external scene event is associated with the user's internal state identifier to determine the current decision context; wherein, when the external scene event is a predefined first type of security-related event and the user's internal state identifier is a predefined first type of high-risk state, the decision context is determined to be a security priority mode.

[0011] Based on the decision context, the user intent data is adaptively filtered or adjusted to generate the final warning trigger command; wherein, in the security priority mode, request-type voice commands that are not related to security are ignored or delayed, and warning trigger commands of the security prompt or alarm type are generated first.

[0012] The output module is used to output corresponding prompt information according to the warning trigger command.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0014] 1. This invention overcomes the shortcomings of existing tour guide systems that make isolated judgments on environmental risks and user status. The system can identify specific external security events and simultaneously assess whether the user is in a high-risk internal state, achieving a more accurate and realistic perception of threats to personal safety.

[0015] 2. This invention solves the core problem of existing systems lacking priority management in information push and command processing. When the system determines that it is in a safety priority mode, it can automatically adaptively filter or adjust non-critical voice commands initiated by the user. In emergency situations, it can prioritize ensuring the smooth flow of the safety channel, thereby ensuring that critical safety prompts or alarm commands can be delivered to the user without delay or interruption, improving the effectiveness and reliability of the system's intervention at critical moments.

[0016] 3. This invention not only passively responds to environmental changes or user requests, but also proactively manages its resources and output strategies within specific decision-making contexts. This proactive, context-aware interactive management mode enhances the overall integrity and intelligence of safety assurance in the ice and snow tourism environment. Attached Figure Description

[0017] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0018] Figure 1 This is a flowchart of the steps of the present invention;

[0019] Figure 2 This is a flowchart of the workflow processing steps of the present invention;

[0020] Figure 3 This is a flowchart illustrating the process of identifying a user's intrinsic state identifier according to the present invention.

[0021] Figure 4 This is a flowchart illustrating the decision-making context of the present invention. Detailed Implementation

[0022] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0023] In existing technologies, safety early warning systems for ice and snow tourism areas largely rely on independent environmental monitoring or user location tracking, making it difficult to achieve accurate risk perception and proactive intervention. Traditional methods often result in single warning messages being easily overlooked when users are focused on guided tours, and they fail to differentiate between individual user states, leading to low warning efficiency. Existing systems cannot correlate external environmental risks with users' internal behavioral states in real time, especially when sudden safety incidents coexist with high-risk user behavior. This makes it difficult to dynamically adjust service priorities and meet the safety requirements of complex tourism scenarios.

[0024] To address the aforementioned issues, the research revealed a risk coupling pattern between external scene events and user behavior states. By establishing a multimodal decision context, intelligent allocation of early warning priorities can be achieved. Further findings indicate that environmental soundprints are highly sensitive to identifying sudden events, while user behavior states show good specificity in individual risk assessment. Therefore, a method for dynamically filtering user requests based on decision context is proposed. Through scenario validation, the correlation between security events and risk states is incorporated into the instruction generation logic, forming a closed-loop early warning system.

[0025] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0026] Example:

[0027] like Figure 1 and Figure 2 As shown, the ice and snow tourism cultural and creative product guidance and anomaly early warning system based on behavior status monitoring includes:

[0028] The data acquisition module is used to acquire environmental soundprint feature data, user behavior status data, and user intent data.

[0029] The multimodal decision module is used to execute:

[0030] External scene events are identified based on environmental soundprint feature data, and user internal state identifiers are identified based on user behavior state data.

[0031] The decision context is determined by associating external scene events with the user's internal state identifier. Specifically, when the external scene event is a predefined first type of security-related event and the user's internal state identifier is a predefined first type of high-risk state, the decision context is determined to be the security priority mode.

[0032] Based on the decision context, user intent data is adaptively filtered or adjusted to generate the final warning trigger command; in the safety priority mode, request-type voice commands that are not related to safety are ignored or delayed, and warning trigger commands of the safety prompt or alarm type are generated first.

[0033] The output module is used to output corresponding prompt information based on the warning trigger command.

[0034] The working principle of this application is as follows: The data acquisition module consists of a microphone array, an inertial measurement unit (IMU, including accelerometer and gyroscope), a GPS / BeiDou positioning chip, and optional heart rate sensor and voice pickup microphone deployed on the user terminal (such as smart glasses, helmet, or handheld device). These sensors work together to collect raw environmental sound signals, user posture and movement trajectory, real-time geographical location, physiological index data, and user voice commands. After preliminary signal conditioning and analog-to-digital conversion, environmental soundprint feature data, user behavior state data, and user intent data that the system can process are formed.

[0035] The multimodal decision-making module is manifested in hardware as a microprocessor or dedicated processing chip (such as an ARM architecture processor or FPGA) in the system's main control circuit, with its decision logic software embedded internally. First, the input multi-source data is processed in parallel:

[0036] On the one hand, by performing time-frequency domain analysis on environmental acoustic signature data (e.g., extracting Mel frequency cepstral coefficients MFCC) and performing pattern matching with a pre-stored acoustic event sample library, external scene events such as snow and ice collapses and strong winds can be identified.

[0037] On the other hand, by extracting and classifying features from user behavior state data (for example, using posture data to determine whether the user is slipping, running, or approaching a dangerous area through a preset algorithm, or analyzing the movement trajectory), an internal user state identifier representing the user's current safety status can be generated.

[0038] The multimodal decision-making module performs a crucial association determination step: it logically associates identified external scene events with the user's internal state identifiers. For example, when the system identifies a safety-related event such as "ice and snow collapse," and simultaneously detects that a user is in a high-risk state of "facing away from the sound source and remaining stationary," the multimodal decision-making module determines the current decision context to be "safety-first mode" based on preset logical rules. In this safety-first mode, the multimodal decision-making module adaptively filters simultaneously input user intent data (e.g., semantic commands converted through speech recognition technology). Specifically, its internal scheduling algorithm prioritizes suspending or delaying request-type voice commands unrelated to current safety, such as "querying the history of a certain scenic spot," and immediately calls the safety warning subroutine to prioritize generating safety prompts or alarm trigger commands containing specific hazard descriptions and avoidance suggestions.

[0039] The output module outputs corresponding prompts based on the received warning trigger command. The output module hardware includes a display screen, a speaker, or a haptic feedback device (such as a vibration motor). For example, it can play a clear voice alarm through a speaker, display a bright flashing warning image on the screen, or drive the device to generate strong vibrations, using multiple sensory channels to ensure that key information can be effectively conveyed to the user.

[0040] As the core analysis module of the system, the multimodal decision-making module realizes the transformation from isolated judgment of the environment or user to situational awareness of the precise target of "specific user in a specific environment" by real-time correlation and fusion analysis of external environmental risks and individual user status, thus providing the initial decision basis for the intelligent response of the entire system.

[0041] As a system-level resource scheduling and instruction management mechanism, the security priority mode dynamically restructures the system's task priorities and interaction logic upon receiving a high-risk decision context. This mode ensures that, in critical situations, the system's output channels and processing resources are forcibly and preferentially allocated to security alert tasks, thus providing an underlying mechanism guarantee for the zero-delay delivery of high-risk alarm information.

[0042] Through the above workflow, this application deeply integrates the originally isolated environmental monitoring and user behavior analysis, enabling the tour guide system to intelligently identify users in high-risk situations and dynamically adjust their interaction strategies. This ensures that in critical situations, life-safety-related early warning information receives the highest processing and output priority, thereby improving the initiative and reliability of safety assurance for ice and snow tourism.

[0043] The system also has a pre-defined strategy for handling decoupled cases:

[0044] When a predefined first-category safety-related event is identified, but the user's internal state is identified as safe or low-risk, the system determines the decision context to be environmental alarm mode. In this mode, the system will not interrupt the current normal guided tour service (such as cultural and creative product explanations), but will generate and output a non-mandatory environmental safety warning. This environmental safety warning has lower priority than warnings in the safety-first mode; for example, it may flash an icon in the corner of the display screen or softly announce "Please note the potential risk of ice ahead" via voice, to avoid unnecessary interference or panic for users not in direct danger.

[0045] When a user's internal state is identified as a predefined first-class high-risk state, but no related external security event is detected: the system determines the decision context to be user behavior alert mode. In this mode, the system will generate an alert based on the user's own behavior, with an intensity between normal navigation and security priority warning.

[0046] For example, if the system detects that a user's movement trajectory is approaching a danger zone, it will output a voice or visual prompt saying "You are approaching a danger zone, please be careful." If it detects an abnormal user posture (such as an impending slip), it will output an immediate warning saying "An imbalance has been detected, please be careful!" When there are no external safety events and the user is in a safe state: the system maintains normal navigation mode, prioritizing and executing the user's routine requests such as cultural and creative information inquiries and route navigation.

[0047] Through the aforementioned multi-mode decision-making logic, the system achieves a refined and tiered response to the overall risk situation of ice and snow tourism. The safety priority mode, as the highest level of response, is triggered only in emergency situations involving risk coupling, ensuring the mandatory and priority nature of key warnings. The environmental alarm mode and user behavior alarm mode serve as important supplements, respectively used to prevent potential environmental threats and correct dangerous user behaviors, together constructing a multi-layered and three-dimensional safety protection system.

[0048] This application further proposes that, when identifying external scene events based on environmental voiceprint feature data, it also includes:

[0049] The system employs signal processing techniques, such as calculating the time difference or phase difference between arrivals at different microphones, to determine the direction of the sound source, i.e., the location of the first sound source. Simultaneously, it performs time-frequency analysis on the sound signal (e.g., extracting features such as Mel-frequency cepstral coefficients) and uses a pre-defined classifier to identify the sound type, thereby determining the type of the first acoustic event.

[0050] To achieve high-precision event recognition, the system relies on a pre-generated scene voiceprint feature map library stored locally on the device or on a cloud server. The training process for this library is as follows: Before system deployment, a large number of standard sound samples of target safety events, including ice collapses, ice surface cracks, and regional strong winds, are collected in typical ice and snow tourism environments (such as glaciers, canyons, frozen lake surfaces, and windy areas), while simultaneously recording their actual locations. Using these sample data labeled with event type and location information, a deep learning model (e.g., a convolutional neural network or a temporal classification model) is trained in a supervised manner. Training conditions include using the Adam optimizer, setting the initial learning rate to 0.001, using cross-entropy as the loss function, and terminating training when the cumulative loss function value no longer significantly decreases over several consecutive training epochs. The final result is a recognition model that can robustly correlate acoustic features with event type and location.

[0051] During actual operation, the system extracts the first acoustic event type and the first sound source location from the environmental acoustic signature data. It then matches these with the corresponding feature entries from a pre-stored scene acoustic signature feature mapping library. This matching identifies the first scene event identifier corresponding to the first acoustic event type and the first sound source location. The predefined first category of safety-related events includes at least: snow and ice collapse sounds, ice surface cracking sounds, and regional strong wind sounds.

[0052] For example, when the system identifies an acoustic event type classified as "high-frequency cracking sound" and its sound source location is determined to be "30 degrees directly in front of the user," it can match the corresponding first-scene event identifier from the mapping library, such as "ice cracking directly in front." In this process, predefined safety event thresholds, such as the threshold used to determine that the spectral energy of ice and snow collapse sound is concentrated in a specific low-frequency range, are pre-set based on the statistical analysis results of a large number of previously measured samples, ensuring the accuracy of the judgment.

[0053] Through the aforementioned technical solution, this application enables the system to transcend simple "noise" judgment in perceiving external risks, reaching a level of qualitative and localization of risk events. By accurately identifying event types, the system can distinguish between threats of different natures (such as snow avalanches and strong winds); and by combining sound source location information, it provides crucial spatial context for further assessing the specific locational relationship between the event and the user, as well as for subsequent collaborative early warnings, thereby significantly improving the accuracy of risk situation assessment and the rationality of subsequent decisions.

[0054] like Figure 3 As shown, this application further proposes that after acquiring the raw signals through the inertial measurement unit, positioning chip, and physiological sensors in the data acquisition module, the system needs to perform refined identification of the user's intrinsic state identifiers. When identifying the user's intrinsic state identifiers based on user behavior state data, the system also includes:

[0055] The system extracts user posture data, user motion trajectory data, and physiological index data from user behavior status data. Specifically, the system filters and calculates the posture of the raw signals from the inertial measurement unit to obtain user posture data describing the angles and movements of various parts of the user's body; it smooths the trajectory and calculates the velocity from the latitude and longitude information from the positioning chip to form user motion trajectory data; and it denoises and extracts features from the raw physiological signals collected by the heart rate sensor and other sensors to obtain key physiological index data such as heart rate and its variability.

[0056] To quantify risk assessment, the system calculates risk coefficients in three dimensions in parallel.

[0057] First, user pose data is matched with predefined typical high-risk poses to calculate the first risk coefficient. In calculating the first risk coefficient In this system development phase, a predefined library of typical high-risk postures is constructed as follows: During system development, a large amount of sensor data is collected from volunteers simulating high-risk postures such as slipping, instability, and curling up in simulated snow and ice environments (e.g., cryogenic laboratories, snow and ice sports training grounds). Using this labeled data, a posture classification model based on dynamic time warping or deep learning is trained. Training conditions typically include using a negative log-likelihood loss function and a stochastic gradient descent optimizer, with a learning rate of 0.01 and a batch size of 32. Training continues until the model's accuracy on the reserved validation set stabilizes. In practical applications, the system inputs real-time posture data into the model. The output probability of matching various high-risk postures, after normalization, is quantified as the first risk coefficient. .

[0058] Secondly, the user's movement trajectory data is compared with the preset safe activity area boundaries imported from the scenic area's geographic information system and stored in the device to calculate the second risk coefficient. In calculating the second risk coefficient, this comparison not only determines whether the user's current location has crossed the boundary, but also predicts whether their movement trajectory in the short term (e.g., the next 10 seconds) is likely to enter the danger zone. Second Risk Coefficient The calculation can be quantified using the following formula:

[0059] ;

[0060] in, It is an indicator function that takes the value 1 if the user has gone out of bounds, and 0 otherwise.

[0061] The straight-line distance from the user's location to the danger boundary. This represents the Euclidean distance (in meters) from the user's current location to the nearest danger boundary.

[0062] and These are preset weighting coefficients used to balance the risk level of boundary crossing behavior with that of adjacent boundaries. Their specific values ​​(e.g., α=0.7, β=0.3) are determined by statistical analysis of location factors in historical accident data.

[0063] Based on physiological indicator data, a third risk coefficient is generated to determine whether the user has a stress response. For example, the system continuously monitors a user's heart rate variability. If the variability is consistently lower than a preset percentage threshold compared to the user's resting baseline (e.g., a decrease of more than 20% lasting for more than 5 seconds; this threshold is derived from commonly used clinical indicators of psychological stress in sports medicine), a stress state is identified. Third risk factor. Weighted values ​​can be assigned based on the number and degree of deviation of such abnormal physiological indicators.

[0064] For the first risk coefficient Second risk factor and the third risk coefficient The system performs fusion processing and determines the user's intrinsic state identifier based on the fusion result; among which, the predefined first type of high-risk state corresponds to the fusion result exceeding the preset state risk threshold.

[0065] In a specific embodiment, a feasible fusion method is weighted summation, which fuses risky results. The calculation formula is as follows:

[0066] ;

[0067] in, It is a preset weight that satisfies The specific allocation of these weights (e.g., =0.4, =0.4, =0.2) can be calibrated based on the contribution of different risk dimensions in historical accident data.

[0068] The system will integrate risk outcomes. With a preset state risk threshold (Based on experimental determination, ranging from 0.5 to 0.8, e.g., 0.65) for comparison. This preset state risk threshold. This was determined through repeated adjustments on the validation set during the model training phase, aiming to maximize the accuracy of risk identification while minimizing the false positive rate. When the risk results are fused... Exceeding the preset risk threshold When this happens, the system determines that the user's internal state is a predefined first-class high-risk state.

[0069] Through the aforementioned multi-dimensional and quantitative risk assessment and fusion mechanism, this application overcomes the one-sidedness and uncertainty of judgment based on a single data source. By comprehensively considering the user's real-time posture, spatial location, and physiological reactions, the assessment of user risk becomes more comprehensive and reliable, providing a solid and quantitative data foundation for the accurate triggering of decision contexts (such as the safety priority mode), and improving the accuracy and credibility of the entire system's early warning.

[0070] like Figure 4 As shown, this application further proposes that after the system identifies external scene events (such as "ice breaking directly in front") and user internal state identifiers (such as "high-risk state") respectively, the step of associating the identified external scene events with user internal state identifiers to determine the current decision context also includes:

[0071] Based on the event identifier of the first scenario, the geographical impact range and event hazard level corresponding to the first scenario event are retrieved from the pre-trained event impact range model. The training data for the event impact range model comes from a historical snow and ice disaster case database, geographic information system data, and fluid mechanics or structural mechanics simulation results. For example, for an "ice surface cracking" event, its geographical impact range can be modeled as an area centered on the crack point and spreading in an irregular polygon along the possible direction of the ice surface crack; while for "snow and ice collapse," its impact range may be a fan-shaped area considering slope and wind direction. The event impact range model is constructed using a gradient boosting decision tree algorithm. The training dataset contains 500 historical snow and ice disaster cases. Each data set includes event type (enumerated value), wind speed (m / s), temperature (°C), and terrain slope (°) as features, and the coordinates (latitude and longitude) of the vertices of the geographical impact range polygon labeled by experts as labels. The model was trained using the Scikit-learn library, with a learning rate of 0.1, a maximum decision tree depth of 6, a subsampling ratio of 0.8, and 100 training iterations until the loss function converged.

[0072] Simultaneously, the system performs short-term trajectory prediction processing on the parsed user motion trajectory data to obtain the user's estimated future location. Trajectory prediction processing can employ linear extrapolation or more complex Kalman filtering algorithms. Its core principle is to combine the user's current position, instantaneous velocity, and direction of movement to predict their estimated future location within a certain time period (e.g., the next T seconds, where T is typically set to an integer between 5 and 15 seconds). This predicted location is a region with a probability distribution, rather than a single point.

[0073] The key step in association determination lies in performing dynamic spatial and risk logic calculations. The system determines whether the user's estimated future location falls within the geographical impact range retrieved from the event impact range model. Simultaneously, it compares the event's hazard level with a preset hazard threshold. This preset hazard threshold is a level value based on safety regulations and the severity of historical incidents; for example, in a hazard level system from 1 (low risk) to 5 (extremely high risk), this threshold can be set to 3 (medium risk and above). The decision context is set to safety-first mode only if both conditions are met simultaneously: the user's estimated future location is within the geographical impact range, and the event hazard level exceeds the preset hazard threshold.

[0074] This application introduces an event impact range model and user trajectory prediction, enabling the system to proactively determine the potential intersection between external environmental threats and the user's future state. This allows the system to trigger the highest level of safety response before the danger actually affects the user. This prediction-based correlation judgment mechanism extends the system's warning lead time and enhances its proactive safety protection capabilities in dynamically changing icy and snowy environments.

[0075] This application further proposes that the step of adaptively filtering or adjusting user intent data based on the decision context also includes:

[0076] Input modality recognition is performed on user intent data to obtain one or more of the following: user voice command data, user gesture interaction data, and user text input data.

[0077] By analyzing the characteristics of the data stream, the system can distinguish between different modalities, such as user voice command data (originating from the microphone), user gesture interaction data (originating from the camera or IMU), and user text input data (originating from the touchscreen or keyboard). In safety-first mode, the system prioritizes processing voice commands that are most likely to consume the user's hearing and attention.

[0078] Specifically, the system inputs user voice command data into a pre-defined command semantic parsing model. This command semantic parsing model is a classification and understanding model based on natural language processing technology, and its training process is as follows: During the system development phase, a large amount of user voice command text in ice and snow tourism scenarios is collected, covering various types such as "cultural and creative information query," "path navigation request," and safety-related commands. These texts are manually annotated, with annotations including command type and command content priority. Command content priority can be graded according to the urgency of the command and its impact on user safety. For example, on a priority scale from 1 (low) to 5 (high), "querying historical allusions" might be labeled as 1, "requesting navigation" as 2, and "asking for help" as 5.

[0079] Using this labeled data, a pre-trained model such as BERT was fine-tuned. Training conditions included using the cross-entropy loss function, the AdamW optimizer, a learning rate of 2e-5, a batch size of 16, and training for 2 to 3 epochs until the model's classification accuracy on the validation set stabilized. This instruction semantic parsing model can simultaneously output the instruction type and corresponding priority value from the text converted from real-time speech recognition.

[0080] The system compares the priority of the instructions identified by the model with a preset security priority threshold. The security priority threshold is a fixed priority value ranging from 1 to 5, preferably set to 3, based on the principle that all non-security instructions of medium to low urgency must be effectively managed. The system also identifies the instruction type and priority of the user's voice instruction data based on a preset instruction semantic parsing model.

[0081] If the instruction type is a cultural and creative information query or a path navigation request, and the instruction content priority is lower than the preset security priority threshold, the system will not execute the instruction immediately. Instead, it will initiate an adaptive adjustment process, which manifests in two ways: delayed processing or replacement with a security confirmation prompt. Delayed processing involves temporarily storing the instruction in a low-priority task queue until the system exits the security priority mode; or replacement processing involves immediately interrupting any potential current output and generating a security confirmation prompt for the user. The security confirmation prompt requests the user's confirmation on whether they still need to execute the original instruction after receiving the prompt.

[0082] The prompt is usually in voice format, such as: "Danger detected on ice ahead. Would you like to review the evacuation guidelines before we search for attraction information?" The system ensures that safety information is delivered mandatoryly, while preserving the user's right to persist with their original request even after confirming certain risks, thus achieving a balance between mandatory safety measures and user autonomy.

[0083] Through the above technical solution, this application achieves a transformation from passively responding to commands to actively managing interactions. When a high-risk situation is identified, the system can intelligently intervene and guide the human-computer interaction process. By delaying or reconfirming non-urgent requests, it forcibly opens a channel for the transmission of critical security information, thereby effectively avoiding security risks that may arise in critical moments due to users being distracted by secondary matters or system resources being occupied by non-critical tasks, ensuring the timeliness and effectiveness of life safety warnings.

[0084] This application further proposes that the system also includes a historical behavior analysis module that communicates with the multimodal decision-making module, enabling continuous self-optimization and personalized adaptation of the system. The historical behavior analysis module is used for:

[0085] During system operation, historical data is continuously acquired and stored. This historical data includes users' historical behavioral status data and corresponding historical decision contexts within a historical period. This historical data is updated on a rolling basis with a configurable historical period (e.g., the most recent 30 days) as the time window. Its content mainly includes users' historical behavioral status data in past periods (such as time-series records of posture, trajectory, and physiological indicators) and the historical decision contexts triggered at the corresponding moments (such as "safety priority mode" or "normal navigation mode").

[0086] Based on this historical data with time-series and contextual labels, a user's specific risk behavior profile is generated using a behavioral pattern learning algorithm. The behavioral pattern learning algorithm can be implemented using unsupervised clustering methods or supervised sequence models. For example, density-based clustering algorithms (such as DBSCAN) can be used to perform cluster analysis on the user's behavioral state data sequence before triggering the "safety priority mode."

[0087] Key parameters of behavior pattern learning algorithms include the radius (eps) defining the neighborhood range and the minimum number of points defining the core object. The method involves conducting multiple clustering experiments on historical behavioral data to determine stable and interpretable cluster results. The training process involves continuously clustering and modeling these historical data to identify unique, recurring behavioral patterns that lead to high-risk states, such as "lingering in a specific area for extended periods to take photos" or "having a stable heart rate but frequently moving near boundaries." These patterns collectively constitute a specific risk behavior profile of the user, a dynamically updated set of user behavior feature vectors.

[0088] This specific risk behavior profile is fed back to the multimodal decision-making module in real time to correct the judgment of the user's intrinsic state. During the multimodal decision-making module's judgment of the user's intrinsic state, this profile serves as an important correction factor. Specifically, if the user behavior state data collected in real time by the system does not reach the high-risk threshold in the general model, but if it highly matches the user's specific high-risk pattern in the profile, the multimodal decision-making module will adjust its fused risk result accordingly. The numerical value. This correction can be achieved through a weighted term based on pattern matching degree, resulting in a corrected fusion risk outcome. The formula is as follows:

[0089] ;

[0090] in, This is the initial result of the fusion risk;

[0091] This represents the maximum matching degree between the real-time behavior sequence and a specific pattern in the user's risk behavior profile, with a value range of [0, 1].

[0092] It is a preset profile correction coefficient used to control the degree of influence of personalized profiles on the general model. Its value is usually set between 0.1 and 0.3. This range is determined based on the premise that the system does not rely too much on personalized data and cause false alarms. The value that can most effectively warn of personalized risks in advance is selected through A / B testing.

[0093] Revised fusion risk results It is then used for the final determination of the user's internal state identifier.

[0094] By introducing a historical behavior analysis module and establishing the aforementioned feedback correction mechanism, the system can identify the unique risks that specific users may face due to their inherent behavioral patterns earlier and more accurately. Thus, it can provide forward-looking warnings based on personalized profiles before the general risk model triggers an alarm, making the system's risk perception capabilities more in-depth and targeted, and improving the accuracy and intelligence level of overall security protection.

[0095] This application further proposes that the system also includes a group collaborative early warning module, which is connected to the system's main processor in hardware via the device's wireless communication unit (such as 5G, Wi-Fi, or an LPWAN module specifically designed for tourist attractions).

[0096] Once the modal decision-making module determines that the decision context is in a safety-first mode and generates an early warning trigger command, the group collaborative early warning module is activated. The primary task is to determine the propagation range of the alert information, i.e., the geographical area covered by the target. The definition of the geographical area covered by the target is not fixed but dynamically calculated based on real-time location information from user behavior state data.

[0097] One specific implementation method is to use the current location of the user who detected the risk as the center, with a dynamic radius... Define a circular coverage area. This dynamic radius... The calculation can take into account environmental factors and risk types, and the calculation formula is as follows:

[0098] ;

[0099] in, It is the basic warning radius, which is preset according to the type of risk event. For example, it is set to 50 meters for "ice surface cracking" and 150 meters for "ice and snow collapse". These values ​​are derived from the physical model estimation of the shock wave range of similar events and the analysis of historical accidents.

[0100] This refers to the speed of risk propagation. For sound-related events (such as the sound of a collapse), the speed of sound (approximately 340 m / s) can be used. For the propagation of ice surface cracks, it can be set to 5-10 m / s based on experience.

[0101] This is the preset system response time, including the total time for information processing, sending, and receiving terminal alarms to take effect, typically set between 2 and 5 seconds. Using this formula, the system can dynamically define a physical space that effectively covers the impact of risks within the response time.

[0102] After determining the geographical area covered by the target, the group collaborative early warning module sends a collaborative early warning trigger command to other guide system terminals located within this area through the communication network deployed in the scenic area. This collaborative early warning trigger command is a structured data packet, which, in addition to containing the core risk event identifier, location, and hazard level, may also include the temporary identification code of the sending terminal. Simultaneously, the group collaborative early warning module also receives collaborative early warning information from other guide systems. To avoid information overload and ensure the accuracy of alarms, the module fuses multiple received early warning messages. For example, a weighted fusion method based on geographical location and credibility can be used: for multiple early warnings received within a short period regarding the same type of risk, the system calculates a fused risk confidence level and performs spatial cross-validation on data from different locations and sources, ultimately generating an integrated, more robust, comprehensive early warning message.

[0103] By introducing a collaborative early warning module, this application achieves a leap from independent single-point perception to networked group perception. It expands the perception boundary of individual users, enabling any localized risk identified by any terminal to be rapidly transformed into an early warning for the entire potentially affected group. This collaborative mechanism, based on dynamic geographic range calculation and multi-source information fusion, overcomes the limitations of a single user's field of vision and the false alarms of a single sensor, thereby constructing an invisible, real-time, interconnected safety net within the scenic area, enhancing the entire tourist area's response speed to sudden risks and its overall safety assurance capabilities.

[0104] Based on the historical behavior analysis module and the group collaborative early warning module using individual risk profiles, the system expands and deepens protective capabilities at both the individual and group levels. At the individual level, by incorporating historical behavioral pattern data to continuously refine real-time risk assessments, the system can anticipate the unique risks of specific users, triggering more proactive early warnings. At the group level, by intelligently disseminating risk events detected at a single point to relevant user groups based on event type, scope of impact, and source identity (e.g., tour guide), the system achieves networked sharing of security information from point to area, constructing a regional collaborative protection network.

[0105] This application further proposes that, during the instruction sending process of the group collaborative early warning module, the system further introduces an intelligent scheduling strategy based on user identity. When sending collaborative early warning trigger instructions to other navigation system terminals located within the target coverage geographical area, the system also includes:

[0106] First, user identification and tour group identification are extracted from user behavior status data. This identification information is bound and stored when the user logs in or the device is initialized. For example, the identification of an ordinary tourist may be "TOURIST", while a tour guide or leader certified by the scenic spot has a special identification of "GUIDE". At the same time, their device records the unique number of the tour group they are in charge of.

[0107] The core logic of the identity verification is as follows: if a user's identity indicates they are a tour guide or leader, the system will automatically trigger a higher-level alert protocol. This is specifically reflected in two aspects. First, the system will increase the priority of collaborative alert triggering commands. The command data packet contains a "transmission priority" field, whose value is typically within a preset range, such as from 1 (lowest) to 5 (highest). For alerts triggered by ordinary tourists, this priority might be set to 3 by default; while for alerts triggered by tour guides, this value will be increased to 4 or 5. This priority setting directly affects the message scheduling order in the scenic area's communication network queue, ensuring that critical alerts are forwarded first, reducing transmission delays. The decision threshold for this priority increase is clear: the increase is triggered only when the identity matches a preset list of tour guide or leader roles.

[0108] Secondly, the system dynamically expands the target coverage area. The base coverage area is based on a dynamic radius. The defined circular area is calculated based on the physical impact of a single risk event. Once the tour guide's identity is identified, the coverage area is no longer limited to this physical model but is expanded to the entire activity range of the tour group associated with the tour guide's tour group identifier. This activity range is a predefined geographic polygon region, whose boundary information comes from the circumscribed convex hull formed by the pre-submitted itinerary plan from the travel agency or dynamically calculated via a real-time positioning system, representing the locations of all members of the tour group within a recent period (e.g., within 15 minutes). The system retrieves the corresponding geographic boundary data from the pre-stored activity range database based on the extracted tour group identifier and uses it as the new, expanded target coverage geographic area.

[0109] By implementing the aforementioned identity-based early warning optimization mechanism, this application achieves hierarchical scheduling and precise allocation of early warning resources. It recognizes and utilizes the tour guide's role as organizer and safety officer within the tour group. When an early warning is initiated by the tour guide's terminal, the system ensures timely information transmission by prioritizing the instruction and expands the warning scope to the entire group's activity area, ensuring that all group members, regardless of their individual location, are forcibly included in the warning coverage. This achieves more efficient and comprehensive safety protection, improving the system's collaborative efficiency and reliability in managing group tourism risks.

[0110] This application further proposes that when the output module outputs the corresponding prompt information according to the warning trigger command, the system further executes a key enhancement function, which aims to transform the abstract alarm into a concrete risk avoidance guide.

[0111] The system first identifies the first scene event identifier (such as "ice breaking" or "regional strong winds") based on environmental soundprint feature data and the user's real-time location (GPS / BeiDou coordinates) parsed from user behavior status data. The system then queries a pre-stored emergency resource geographic information database to match the location of the nearest emergency resource point.

[0112] The emergency resource geographic information database is a structured database whose content is entered by the scenic area management department before system deployment and is updated regularly. It includes the latitude and longitude coordinates, types, and corresponding risk events of various emergency resource points. For example, for an "ice surface breakage" event, the relevant emergency resource point types may include "lifebuoy drop points," "emergency heating stations," and "scenic area rescue stations"; while for "regional strong winds," it may correspond to "emergency shelters."

[0113] The matching process is not simply about finding the nearest point, but rather about intelligent filtering based on the association between event type and resource function. The system first filters out all emergency resource point types corresponding to the current first scenario event identifier from the database, forming a candidate set. Then, the system calculates the Euclidean distance between the user's real-time location and the location of each emergency resource point in this candidate set.

[0114] In this embodiment, an optimized matching strategy is to define a comprehensive priority score. To sort them, the formula is as follows:

[0115] ;

[0116] in, Indicates the distance (in meters) from the user to the emergency resource point;

[0117] This is a type weighting coefficient, whose value is pre-set within a range, such as 1.0 to 3.0, based on the importance of the emergency resource point to the current specific risk event. For example, for the "ice breakage" event, the "lifebuoy drop point"... The highest priority score could be set to 3.0, while "emergency heating station" could be set to 2.0. This is determined by calculating the overall priority score for each candidate point. The scores are sorted in descending order, and the system can match the nearest emergency resource point location that takes into account both functional relevance and spatial proximity. Here, "nearest" refers to the weighted optimal solution, rather than simply the geometric nearest.

[0118] Once the target emergency resource point is identified, the system will generate a prompt message that includes not only a risk description but also a navigation path to the nearest emergency resource point. This navigation path information is generated in real time based on scenic area map data by calling the system's built-in path planning algorithm. It can be overlaid on the user's navigation interface map as a highlighted arrow or output as a concise voice command (such as "Please proceed east along the current road for 100 meters, then turn left to reach the emergency shelter").

[0119] Through the aforementioned technical solution, this application achieves a functional enhancement from risk warning to risk avoidance guidance. While issuing warnings, it provides users with clear, immediate, and actionable plans. By generating navigation prompts that match specific risk types and point to clearly defined risk avoidance resources, it effectively reduces user panic and decision-making confusion in emergency situations, guiding users to quickly and orderly evacuate to safe areas or obtain critical rescue supplies. This enhances the practicality of the warning information and the overall risk avoidance guidance effectiveness of the system.

[0120] To ensure reliable system operation even in the event of partial hardware failure, this application further introduces a device status monitoring module directly connected to the data acquisition module. This module can be integrated into the main processor and continuously monitors the operational status data of various sensors in the data acquisition module through specific drivers and diagnostic circuits. This status data includes, but is not limited to: microphone signal level and signal-to-noise ratio, inertial measurement unit (IMU) self-test output and data update rate, positioning chip satellite lock status, and physiological sensor signal quality index. The module internally presets parameter ranges for normal operation of various sensors; for example, the microphone signal-to-noise ratio should be higher than 20dB, and GPS positioning data should not exhibit prolonged periods of fixed value drift.

[0121] The monitoring module analyzes this status data in real time. Its core judgment logic is based on identifying abnormal or missing data from key sensors. "Key sensors" here refer to components indispensable for core safety judgments, such as microphone arrays for identifying ambient sound patterns and IMUs for determining user posture. Anomaly detection is based on preset fault thresholds. For example, if the microphone signal remains below the noise baseline for more than 5 seconds, or if IMU data remains unchanged for 10 consecutive cycles, it is considered abnormal. Once such conditions are met, the monitoring module performs two linked operations: First, it generates a device fault alarm, which is output through the module in a specific form (such as screen icon flashing or periodic slight vibrations) to alert the user that the device's functionality is limited; second, it triggers the multimodal decision module to switch to a degraded decision mode.

[0122] In degraded decision-making mode, the system's decision logic will adaptively adjust to maintain robustness. The multimodal decision module will then continue risk assessment based on the remaining valid sensor data. For example, if the microphone malfunctions, the system will rely solely on user behavior data (posture, trajectory, physiological indicators) for judgment. To compensate for potential perceptual uncertainty due to reduced data sources and to prevent false alarms, the system will synchronously increase a preset risk threshold. Specifically, a preset state risk threshold is used to determine the user's internal state identifier. This will be dynamically adjusted. The adjusted preset risk threshold... The adjustment formula is as follows:

[0123] ;

[0124] in, It is the preset risk threshold in normal mode (e.g., 0.65).

[0125] It is the total number of key sensors;

[0126] This represents the number of key sensors that are still functioning normally.

[0127] It is a preset threshold compensation amount, which is usually in the range of 0.1 to 0.3. This range is determined by system simulation to suppress the false alarm rate caused by sensor deficiency to the greatest extent while ensuring a certain false alarm rate.

[0128] Using this formula, the system uses the adjusted preset state risk threshold in degraded mode. This generates warning trigger commands, making the trigger conditions more stringent.

[0129] By introducing equipment status monitoring and degradation decision-making modes, the system transitions from full-function operation under ideal conditions to functional preservation and robust operation under non-ideal conditions. When some of the system's sensing capabilities are impaired due to hardware failure, it can automatically reduce the "sensitivity" of its judgments by dynamically adjusting its internal decision parameters. This effectively suppresses a large number of unreliable alarms that may be generated due to incomplete data while maintaining safety warning functions as much as possible. This ensures the reliability of the system's output and user experience under non-perfect conditions, and improves the system's usability and robustness throughout its entire lifecycle.

[0130] In complex real-world scenarios, the system may simultaneously identify multiple external events and / or multiple high-risk user states, leading to conflicting decision-making contexts. To ensure the rationality and uniqueness of the system's response, this system introduces a priority adjudication mechanism based on a weighted average of event hazard level and risk coefficient.

[0131] When multiple external scene events are identified simultaneously, the system prioritizes the event with the highest danger level as the primary basis for decision-making. For example, if both "regional strong winds" (preset danger level 3) and "ice breakage directly ahead" (preset danger level 4) are identified at the same time, the system will use the "ice breakage" event, which has a higher danger level, as the dominant event for subsequent correlation determination.

[0132] When user behavior data analysis identifies multiple high-risk factors (such as the first risk factor) Extremely high, and also the second highest risk factor. (Also extremely high), the system identifies the user's internal state after fusion (i.e., the fusion risk result). As the final criterion, the fusion result itself already contains a weighted synthesis of multiple states.

[0133] In some extreme cases, indications from different modalities may conflict. For example, an environmental event may have a high hazard level, but the user's fusion risk results may differ. However, the threshold was not reached. Therefore, the system has set a final decision logic:

[0134] 1. Environmental Event Priority Principle: For predefined, immediately lethal, highest-risk events (e.g., level 5 events such as large-scale snow and ice collapses), as long as the event's hazard level exceeds a preset absolute hazard threshold, the system will forcibly trigger the safety priority mode regardless of the user's current internal status. This principle is based on the logic of "preserving life as the highest priority," prioritizing responses to the most urgent external threats.

[0135] 2. Weighted Fusion Principle: For events that are not at the highest risk level, the system adopts a weighted fusion decision score. This determines the final decision context. The weighted fusion decision score... The calculation formula is as follows:

[0136] ;

[0137] in, The danger level of the identified dominant external scene events (normalized to the 0-1 range).

[0138] The user's integration risk outcome (0-1 range);

[0139] and The preset weighting coefficients, and According to a conservative security strategy, it is usually set to... (For example =0.6, =0.4), giving environmental events a higher decision-making weight.

[0140] If the weighted fusion decision score If the threshold for mode switching is exceeded, the system will enter the safety priority mode; otherwise, it will output an alarm of the corresponding level according to the single-factor logic.

[0141] Through the aforementioned multi-level adjudication mechanism, the system can intelligently handle conflicts of information from multiple sources, ensuring that it can output the most reasonable and reliable security decision under any complex circumstances, thus achieving a complete closed loop in logic.

[0142] The following is a specific implementation example of a snow and ice tourism cultural and creative product guidance and anomaly early warning system based on behavior status monitoring:

[0143] A tourist used a smart tour guide wristband (built-in 4-channel microphone array, 16kHz sampling rate; 6-axis IMU, 100Hz sampling rate; GPS module, 1-meter positioning accuracy; optical heart rate sensor, 1Hz sampling rate) to tour an ice sculpture art area in a northern high-altitude ice and snow scenic area. The wristband collected environmental soundprint data, user behavior data, and voice intent data in real time: the environmental soundprint captured the sound of ice cracking 100 meters away (acoustic event type "high-frequency cracking sound", sound source direction is 30 degrees to the left of directly in front of the user); the user's posture was stationary with their back to the sound source (focused on taking pictures), the trajectory was only 5 meters away from the dangerous boundary of the ice surface, and the heart rate was 15% higher than the resting value; the voice intent was "explain the history of this ice sculpture".

[0144] The system's multimodal decision-making module first processes environmental acoustic data, extracting acoustic event types and locations. After matching with the scene acoustic database, it identifies the event as a Class I safety event (ice surface breakage, danger level 4, normalized Le=0.8). Next, it analyzes user behavior status: the first risk coefficient... =0.7 (matching the high-risk posture of "stillness with back to the sound source"), second risk coefficient =0.6 (trajectory close to danger boundary), third risk coefficient =0.3 (heart rate slightly high), the fusion risk result after fusion is =0.4×0.7+0.4×0.6+0.2×0.3=0.58.

[0145] The system predicts that the user's trajectory will enter the ice crack's impact area within the next 10 seconds (the geographical impact area is a 100-meter circular area centered on the crack point), and the event's danger level exceeds the preset threshold of 3. The decision context is set to safety priority mode. At this time, the user's voice intent is semantically interpreted as "cultural and creative product query" (priority 1), which is below the safety threshold of 3. The system delays processing this request and immediately outputs a warning voice: "Ice crack ahead, high danger level, please immediately evacuate to the emergency shelter 50 meters to the right," and displays a navigation path (to the nearest emergency resource point) on the wristband screen. The group collaborative warning module uses the user's location as the center and a dynamic radius... The system covers an area potentially affected by ice crevasses and sends a coordinated alert to 12 tourists in that area. Tourists who receive the alert will have their wristbands displaying evacuation instructions.

[0146] The weighted fusion adjudication score is calculated using the following formula. ,in =0.6 (Environmental event weight) =0.4 (user state weight), substituting this into the weighted fusion decision score. =0.6×0.8+0.4×0.58=0.712, exceeding the preset mode switching threshold of 0.6, triggering the safety priority mode.

[0147] Upon hearing the warning, tourists immediately stopped taking photos and followed the navigation to a safe area, demonstrating that they did not overlook the danger due to their focus on the cultural and creative product explanations. The system accurately identified their high-risk status (facing away from the sound source) and prioritized sending warnings rather than broadcasting them to all tourists. Group collaboration allowed nearby tourists to take precautions in advance, avoiding potential group dangers. When the wristband microphone malfunctioned (signal-to-noise ratio below 20dB), the device status monitoring module triggered a degradation mode, increasing the preset status risk threshold (from 0.65 to 0.75). It could still determine trajectory risks based on IMU and GPS, maintaining basic warning functions and ensuring safety even if some hardware failed.

[0148] This embodiment clearly demonstrates how the system solves problems such as "isolated judgment of environment and user status" and "lack of early warning priority". Through multimodal fusion and dynamic decision-making, it achieves accurate early warning and proactive intervention in complex ice and snow scenarios, improving the balance between tourist safety and guided tour experience.

[0149] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. An ice and snow tourism creative guide and abnormal warning system based on behavior state monitoring, characterized in that: Comprising: a data acquisition module for acquiring environmental acoustic feature data, user behavior state data, and user intent data; a multi-modal decision-making module for performing: identifying an external scene event based on the environmental acoustic feature data and a user internal state identifier based on the user behavior state data; associating the external scene event with the user internal state identifier to determine a current decision-making context; wherein when the external scene event is a predefined first type of safety-related event and the user internal state identifier is a predefined first type of high-risk state, the decision-making context is determined to be a safety priority mode; according to the decision-making context, adaptively filtering or adjusting the user intent data to generate a final warning trigger instruction; wherein in the safety priority mode, non-safety-related request type voice instructions are ignored or delayed, and safety prompt or alarm type warning trigger instructions are preferentially generated; an output module for outputting corresponding prompt information according to the warning trigger instruction.

2. The behavior state monitoring-based ice and snow tourism cultural and creative guide and abnormality early warning system according to claim 1, characterized in that: When identifying the external scene event based on the environmental acoustic feature data, further comprising: extracting a first acoustic event type and a first sound source direction from the environmental acoustic feature data; matching a first scene event identifier corresponding to the first acoustic event type and the first sound source direction from a pre-stored scene acoustic feature mapping library; wherein the predefined first type of safety-related event at least includes ice and snow collapse sound, ice surface cracking sound, and regional strong wind sound. 3.The ice-snow tourism creative guide and abnormality early warning system based on behavior state monitoring according to claim 2, characterized in that: When identifying the user internal state identifier based on the user behavior state data, further comprising: parsing user posture data, user motion trajectory data, and physiological indicator data from the user behavior state data; matching the user posture data with a predefined typical high-risk posture to calculate a first risk coefficient; comparing the user motion trajectory data with a pre-set safety activity area boundary to calculate a second risk coefficient; determining whether the user has a stress physiological response based on the physiological indicator data to generate a third risk coefficient; fusing the first risk coefficient, the second risk coefficient, and the third risk coefficient, and determining the user internal state identifier based on the fusion result; wherein the predefined first type of high-risk state corresponds to the fusion result exceeding a pre-set state risk threshold.

4. The behavior state monitoring-based ice-snow tourism cultural and creative guide and abnormality early warning system according to claim 3, characterized in that: The step of associating the identified external scene event with the user internal state identifier to determine the current decision-making context further comprises: according to the first scene event identifier, retrieving a geographical influence range and an event danger level corresponding to the first scene event from a pre-trained event influence range model; performing trajectory prediction processing on the user motion trajectory data to obtain a user future estimated position; when the future estimated position is located within the geographical influence range and the event danger level exceeds a pre-set danger threshold, determining the decision-making context to be a safety priority mode.

5. The behavior state monitoring based ice-snow tourism cultural and creative guide and abnormality early warning system according to claim 1, characterized in that: The step of adaptively filtering or adjusting the user intent data according to the decision-making context further comprises: input modality recognition is performed on the user intention data to obtain one or more of user voice instruction data, user gesture interaction data, and user text input data; in the safety priority mode, an instruction type and an instruction content priority in the user voice instruction data are identified based on a preset instruction semantic analysis model; if the instruction type is a creative information query or a path navigation request and the instruction content priority is lower than a preset safety priority threshold, a delay processing or a replacement with an output safety confirmation prompt is performed; the safety confirmation prompt is used to request the user to confirm whether the original instruction still needs to be executed after receiving the prompt information. 6.The ice-snow tourism creative guide and abnormality early warning system based on behavior state monitoring according to claim 1, characterized in that: The system further comprises a historical behavior analysis module in communication connection with the multi-modal decision module, and the historical behavior analysis module is configured to: acquire and store historical data, the historical data comprising historical behavior state data of the user in a historical period and corresponding historical decision context; generate a specific risk behavior portrait of the user based on the historical data through a behavior pattern learning algorithm; feed back the specific risk behavior portrait to the multi-modal decision module for correcting the determination of the internal state identification of the user.

7. The behavior state monitoring-based ice-snow tourism cultural and creative guide and abnormality early warning system according to claim 6, characterized in that: The system further comprises a group coordination early warning module, configured to: after determining that the decision context is in the safety priority mode and generating a warning trigger instruction, determine a target coverage geographic area of the prompt information based on real-time location information in the user behavior state data; send a coordination early warning trigger instruction to other guide system terminals located in the target coverage geographic area; receive coordination early warning information from the other guide systems and output after fusion.

8. The behavior state monitoring-based ice-snow tourism cultural and creative guide and abnormality early warning system according to claim 7, characterized in that: When sending a coordination early warning trigger instruction to other guide system terminals located in the target coverage geographic area, the method further comprises: extracting a user identity and a travel group identity from the user behavior state data; if the user identity indicates a group tour guide or a leader, increasing the priority of the coordination early warning trigger instruction and expanding the target coverage geographic area to the entire travel group activity range. 9.The ice-snow tourism creative guide and abnormality early warning system based on behavior state monitoring according to claim 1, characterized in that: When the output module outputs corresponding prompt information according to the warning trigger instruction, the method further comprises: matching the location of the nearest emergency resource point from a pre-stored emergency resource geographic information database according to the first scene event identification and the real-time location of the user identified from the environmental voiceprint feature data; wherein the prompt information comprises a navigation path to the nearest emergency resource point.

10. The behavior state monitoring based ice-snow tourism cultural and creative guide and abnormality early warning system according to claim 1, characterized in that: The system further comprises a device state monitoring module connected with the data acquisition module, and the device state monitoring module is configured to: monitor the running state data of various sensors in the data acquisition module; if a critical sensor data anomaly or absence is identified, a device fault alarm is generated and the multi-modal decision module is triggered to switch to a degraded decision mode; in the degraded decision mode, the multi-modal decision module generates the warning trigger instruction based on the remaining valid sensor data and increases a preset state risk threshold.

Images